AE-YOLO: An Improved YOLOv7 based on Attention Enhancement for Rail Flaw Detection

Abstract

With the development of rail transit, railway is becoming an increasingly common means of transportation, which puts forward higher requirements for its safety and reliability. Meanwhile, as one of the most important infrastructures, the failure of steel rail will seriously affect the safe and stable operation of the railway, so it is necessary for us to monitor its internal flaws. An improved You Only Look Once Version 7 based on Attention Enhancement is proposed for rail flaw detection. New detection layer is added for small rail flaws detection in the proposed model AE-YOLO. Convolutional block attention module (CBAM) is introduced into the model, which can make detection model pay more attention to important flaws that may affect the rail safety. On the B-scan dataset collected by the big detector train, the AE-YOLO model is tested. Compared with YOLOV7, the proposed model can increase the mean average precision by 4.5%. And recall is also improved by 3.4%, which is an extremely important indicators for rail flaw detection. The case study results indicate the efficiency and superiority of the proposed AE-YOLO model in rail flaw detection task.